Deep Learning COVID-19 Features on CXR using Limited Training Data Sets

Yujin Oh; Sangjoon Park; Jong Chul Ye

Source-linked AI summary

Deep Learning COVID-19 Features on CXR using Limited Training Data Sets

Yujin Oh, Sangjoon Park, Jong Chul Ye

arXiv:2004.05758v2Apr 13, 2020eess.IVcs.CVcs.LGstat.ML

TL;DR

Limited, heterogeneous CXR data complicate deep-learning development for COVID-19 diagnosis and triage. The paper addresses this with a biomarker-motivated patch-based network and probabilistic Grad-CAM, reporting improved sensitivity, low parameter count, and interpretable saliency maps.

Problem
Limited systematic collection of well-curated CXR data makes training neural networks for COVID-19 diagnosis difficult, despite the potential value of CXR for diagnosis and triage.
Method
The paper proposes a patch-based convolutional network motivated by CXR intensity-distribution analysis, using majority voting and probabilistic Grad-CAM for global interpretation.
Results
92.5% sensitivity was reported for COVID-19 and viral cases, while the method used 11.6 M parameters versus COVID-Net’s 116.6 M.
Takeaways & Limitations
The method provides a data-efficient approach with comparative performance and saliency maps reported to correlate with radiological findings, supporting COVID-19 diagnosis and triage applications.

Abstract

from arXiv · show

Under the global pandemic of COVID-19, the use of artificial intelligence to analyze chest X-ray (CXR) image for COVID-19 diagnosis and patient triage is becoming important. Unfortunately, due to the emergent nature of the COVID-19 pandemic, a systematic collection of the CXR data set for deep neural network training is difficult. To address this problem, here we propose a patch-based convolutional neural network approach with a relatively small number of trainable parameters for COVID-19 diagnosis. The proposed method is inspired by our statistical analysis of the potential imaging biomarkers of the CXR radiographs. Experimental results show that our method achieves state-of-the-art performance and provides clinically interpretable saliency maps, which are useful for COVID-19 diagnosis and patient triage.

I. INTRODUCTION

The paper motivates CXR-based COVID-19 diagnosis and triage because RT-PCR has sampling-related limitations, CT can burden radiology services, and CXR sensitivity is comparatively low. It therefore develops a patch-based, data-efficient network with interpretable saliency maps.

Early COVID-19 detection is important for limiting transmission and allocating scarce medical resources.
RT-PCR is the diagnostic standard, but nasopharyngeal and throat-swab results can be affected by sampling errors and low viral load.
CXR shows COVID-19-related radiological abnormalities but has lower reported sensitivity than initial RT-PCR testing.
Improved CXR performance could support COVID-19 identification and patient triage when healthcare systems face resource constraints.
The proposed approach investigates CXR biomarkers and uses patch-based classification with a relatively small network and probabilistic Grad-CAM saliency maps.
The overall framework normalizes CXR data, segments lung areas, classifies local patches, combines patch predictions by majority voting, and produces saliency maps.

A. Segmentation network

The method segments anatomical regions before classification, normalizes heterogeneous CXR data, and uses a local patch-based classifier designed for limited training data. The classification setup also uses a simple backbone and clinically motivated disease labels.

Segmentation network: The segmentation network extracts lung and heart contours from chest radiographs using an extended FC-DenseNet103.
Segmentation network: The segmentation objective is a weighted cross-entropy loss for four semantic classes: background, heart, left lung, and right lung.
Preprocessing: Preprocessing casts image data to float32, equalizes histograms, applies gamma correction with γ = 0.5, and resizes images to 256 × 256.
Classification network: The classifier uses ResNet-18 and ImageNet transfer learning to reduce overfitting and stabilize training with few samples.
Classification network: Images are labeled as normal, bacterial pneumonia, tuberculosis, or viral pneumonia, with COVID-19 grouped with other viral pneumonias.
Classification network: The global baseline resizes masked images to 224 × 224, whereas the proposed local approach randomly crops 224 × 224 patches from 1024 × 1024 lung-region images.

C. Probabilistic Grad-CAM saliency map visualization

The paper adapts Grad-CAM for local patch classification by combining patch-level activation maps with class probabilities and averaging their contributions into a global saliency map.

Grad-CAM: Grad-CAM produces a class-specific saliency map from the final convolutional feature channels, upsampling, and a ReLU operation.
Grad-CAM: The feature-channel weights are determined for the target class before the softmax layer.
Grad-CAM: The global approach uses standard Grad-CAM as its baseline saliency map, normalized to values between 0 and 1.
Probabilistic Grad-CAM: Because local patches receive different COVID-19 scores, the proposed probabilistic Grad-CAM weights patch saliency maps by estimated class probabilities before averaging them.

III. METHOD

The method uses public CXR datasets with segmentation masks for segmentation training and curated, diagnosis-labeled datasets for five-class classification. Pediatric images were excluded to reduce age-related bias in COVID-19 classification.

Segmentation network dataset: Public datasets with ground-truth masks were used to train and validate the segmentation networks.The JSRT/SCR dataset was split into training and validation sets, with the NLM(MC) dataset used for cross-database validation.
Classification dataset: Classification data combined segmentation datasets with additional public COVID-19 resources carrying detailed disease labels.These labels supported separate classification of bacterial and viral pneumonia.
Classification dataset: Pediatric CXR images were excluded because more than 80% of one public dataset was pediatric, creating a risk of age-related feature bias.The exclusion targeted a more unbiased age distribution for evaluating COVID-19 classification.
Classification dataset: The classification dataset was curated into normal, tuberculosis, bacterial pneumonia, viral pneumonia, and COVID-19 pneumonia classes.The collection included 180 radiography images from 118 COVID-19 subjects and additional images from other disease classes.

3) Dataset for comparison with COVID-Net:

The COVID-Net comparison used a separately prepared dataset, while the method analyzed CXR biomarkers and evaluated classification with standard confusion-matrix metrics. Biomarkers included lung morphology, intensity statistics, and cardiothoracic ratio.

Dataset for comparison with COVID-Net: A separate dataset combined COVID-19 image data with the RSNA Pneumonia Detection Challenge dataset for comparison with COVID-Net.The additional comparison used labels divided into three categories because the RSNA dataset provided pneumonia presence rather than detailed disease diagnosis.
Statistical analysis of potential CXR COVID-19 markers: The biomarker analysis examined lung morphology, mean lung intensity, standard deviation of lung intensity, and cardiothoracic ratio.Lung-area intensity measures were derived from segmented lung pixels, while CTR used cardiac and thoracic diameters.
Statistical analysis of potential CXR COVID-19 markers: CTR was treated as a potential abnormality alarm because COVID-19 opacities or consolidation may blur the cardiothoracic boundary.CTR divides maximal transverse cardiac diameter by maximal internal thoracic diameter.
Classification performance metrics: Classification performance used accuracy, precision, recall, F1 score, and specificity calculated from confusion-matrix counts.F1 score served as the early-stopping metric, and overall scores were averaged across classes.

IV. EXPERIMENTAL RESULTS

Experiments evaluated segmentation across databases, morphological and intensity biomarkers, and the role of segmentation masks in classification. Preprocessing improved cross-database segmentation, while lung intensity differed significantly but overlapped across classes.

Segmentation performance on cross-database: Jaccard similarity results showed segmentation accuracy comparable to previous work on the JSRT and NLM(MC) datasets.Table V reports the coefficient for each contour on the validation dataset.
Segmentation performance on cross-database: Jaccard similarity coefficients increased from 0.932 to 0.943 after preprocessing NLM images, with p < 0.001.The result identifies universal intensity normalization as important for cross-database segmentation.
Morphological analysis of lung area: Segmentation masks showed class-dependent morphological differences, including deformation in bacterial cases and partial deformation in viral cases.COVID-19 examples retained fully segmented lung areas despite bi-basal infiltrations, while morphology was considered potentially useful for severe-infection discrimination.
Lung areas intensity: COVID-19 cases had lower mean lung intensity than normal, bacterial, and tuberculosis cases with statistical significance, but the class distributions broadly overlapped.The reported significance levels were p < 0.001 for normal and bacterial cases and p < 0.01 for tuberculosis.

2) Lung areas intensity variance:

Statistical analysis identified lung-area intensity variation as a strong COVID-19-associated feature, while cardiothoracic ratio showed broad class overlap. Patch-level analysis suggested globally distributed intensity changes were more informative than local texture variation.

COVID-19 and viral cases showed significantly higher lung-area intensity variance than other classes (p <0.001 for all).
Excluding AP Supine radiographs produced only minor differences, indicating high intensity variability in COVID-19 and viral classes was invariant to scanning protocol.
COVID-19 cases differed statistically from normal, tuberculosis, and bacterial classes in cardiothoracic ratio, but the distributions broadly overlapped.
Lung-area intensity distribution was identified as potentially effective for diagnosis because it reflects multifocal peripheral and lower-zone consolidation and ground-glass opacities.
Intensity variance is not uniquely specific to COVID-19 because bacterial pneumonia can produce similar distributions through different opacity-spreading patterns.
Inter-patch intensity distributions distinguished the unified COVID-19 and viral class, whereas intra-patch intensity distributions did not differ from normal cases.
COVID-19 and viral cases were combined because all examined marker candidates showed no significant difference between them (p >0.05).

D. Classification performance

The local patch-based classifier outperformed the global approach across reported metrics and produced saliency maps that better represented multifocal COVID-19 lesions. Its COVID-19/viral sensitivity reached 92.5%.

The proposed local patch-based approach performed consistently better than the global approach across all metrics.
92.5% sensitivity was achieved for the combined COVID-19 and viral class.
The confusion matrices compare classification errors for the global and local patch-based approaches.
Saliency examples included COVID-19 patients and normal, bacterial, tuberculosis, and COVID-19 pneumonia cases.
The local probabilistic Grad-CAM visualized multifocal ground-glass opacities and consolidations more effectively than the global Grad-CAM.

V. DISCUSSION

The discussion frames CXR as a potential triage tool rather than a complete replacement for RT-PCR or CT. The proposed workflow would reserve confirmatory testing for patients whose CXR classification remains difficult.

COVID-19-mimicking diseases, including bacterial and other coronavirus pneumonias, must be differentiated during diagnosis.
Global localized intensity variation was identified as a discriminatory COVID-19 CXR factor and was confirmed by saliency maps.
The triage rationale reflects the substantial prevalence of bacterial pneumonia and tuberculosis during the pandemic.
The proposed network groups other viral infections with COVID-19 and excludes normal, bacterial pneumonia, and tuberculosis cases at an early stage.
RT-PCR or CT can then be reserved for patients classified as other virus and COVID-19 cases, whose diagnosis remains difficult by CXR.
The proposed algorithm is presented as a potential triage workflow for COVID-19 patients.

C. Training stability

Patch-based training improved stability under limited data while retaining competitive performance against COVID-Net with far fewer parameters. The study’s dataset was not rigorously curated and came from only one or two sources.

Patch-based training showed no sign of overfitting, whereas the global approach showed significant overfitting despite equal numbers of weight parameters.
91.9% overall accuracy for the proposed method was comparable to 92.4% for COVID-Net on the same dataset.
The comparison with COVID-Net was conducted using the same dataset.
11.6 M parameters were used by the proposed method versus 116.6 M for COVID-Net, approximately one-tenth as many.
The proposed method provided significantly improved sensitivity to COVID-19 cases compared with COVID-Net.
The study was limited by poorly curated CXR data originating from only one or at most two sources, with public COVID-19 images largely gathered from online materials.

F. Segmentation network analysis

The segmentation analysis found that FC-DenseNet103 offered the best complexity–performance trade-off among evaluated segmentation networks, while severe consolidation caused unavoidable degradation. Segmentation masks nevertheless improved classification overall and carried discriminatory morphological information.

Segmentation-network comparison: FC-DenseNet103 consistently outperformed the proposed alternatives across classification metrics and significantly improved performance over the smaller FC-DenseNet67.Its better complexity–performance trade-off motivated its adoption as the segmentation network.
Limited-data robustness: Jaccard similarity remained stable until 50% of the training set was used on the preprocessed NLM(MC) dataset.Without preprocessing, segmentation performance decreased as the training set was reduced.
Segmentation performance: The FC-DenseNet103 segmentation network achieved comparable state-of-the-art performance on normal subjects.
Segmentation limitations: Severe consolidation caused unavoidable segmentation degradation because such images were absent from training.Bacterial pneumonia radiographs with widely spread opacity could produce under-segmented lung regions.
Segmentation as a classification marker: Binary segmentation masks distinguished normal from abnormal images with 86.9% sensitivity, indicating discriminatory morphological information.Under-segmentation also affected disease classification, particularly for bacterial cases.
Effect on classification: Using segmentation masks significantly improved overall classification performance despite under-segmented outliers.The authors identify further training on patient cases with correct labels as an opportunity for improvement.

G. Classification network analysis

The classification analysis evaluated patch size and training-set size for the proposed local patch-based approach. Unlike whole-image classification, the patch-based method maintained robustness as the dataset became smaller, supporting its use with limited data.

Patch-size analysis: 112 × 112 patches produced worse classification results, while 448 × 448 patches offered no benefit over the 224 × 224 patch size.The analysis therefore identified drawbacks to reducing patch size and no improvement from increasing it.
Dataset-size analysis: The whole-image global approach showed a prominent accuracy decrease as dataset size decreased, whereas the proposed local patch-based method showed greater robustness.Dataset-size effects were evaluated because the method was designed for limited-data settings.
Motivation: Systematic collection of large CXR datasets for deep neural-network training was difficult during the emergent pandemic.
Biomarker basis: The method investigated localized intensity variation as a potential discriminatory CXR feature for COVID-19.
Patch-based classification: Multiple randomly located lung patches were classified and combined by majority voting, with probabilistic Grad-CAM providing an interpretable saliency map.The approach was reported to train stably with small datasets and produce saliency maps correlated with radiological findings.